A hybrid linear mixed model-neural network framework for pavement roughness forecasting from vehicle IRI measurements

Abstract

Forecasting pavement deterioration remains challenging due to limited labelled data and the nonlinear dynamics of degradation. This study proposes a hybrid methodology that leverages Linear Mixed Models (LMMs) to encapsulate longitudinal International Roughness Index (IRI) measurements and Neural Networks (NNs) to generalise deterioration patterns across streets. The LMM compresses raw time series into interpretable linear trends and confidence intervals, consistent with the observation that deterioration is approximately linear in its early stages. A sliding-window training scheme is then introduced: LMMs trained on shorter historical spans (1–3 years) provide input parameters to the NN, while the subsequent year’s LMM outcome serves as the supervisory training target. This reduces computational costs relative to direct sequence models while retaining predictive accuracy. Additionally, confidence intervals from the LMM enable unsupervised anomaly detection, flagging segments where new IRI measurements indicate accelerated degradation. The results on multi-year road data from the Port of Antwerp in Belgium demonstrate that the hybrid LMM-NN approach achieves robust 1-year-ahead IRI forecasts and provides early-warning indicators of deterioration without requiring explicit condition labels. The framework offers a scalable, interpretable alternative to purely data-driven models for pavement management applications.